Making Compute a Tradable Commodity: Carmen Li on Building the Data and Markets Layer for AI Infrastructure
Our first guest on Where Capital Meets Compute is Carmen Li, founder and CEO of Silicon Data and CEO of Compute Exchange. The two companies are opposite sides of the same coin. Silicon Data is the data and benchmarking layer, building the indices that let people reference GPU prices the way they reference oil or rates. Compute Exchange is the physical delivery side, a marketplace where buyers procure GPU and token capacity through forward and reserve contracts.
Carmen came from capital markets, starting as an energy options market maker in Chicago hedging oil, then moving to Citi and Bloomberg, where she ran global data integrations. That mix of market maker instinct and data infrastructure discipline runs through everything she builds. The headline milestone: her compute futures, built with CME, launch in October.
The problem: exposure no one could hedge
Carmen started in 2024, when there was still no way to price or hedge compute. Her early engagement was with AI startups buying GPU on demand, where the rate might swing from three dollars to nine and back. More clients meant more exposure to volatility, not better economics. Her framing: a business that cannot lock its input cost cannot price its output with confidence.
American Airlines cannot issue you a ticket price if they cannot hedge their oil prices. It’s just very intuitive. The market needs indices and a way to manage volatility.
A token factory should focus on deploying models efficiently, not guessing whether compute costs three dollars tomorrow. One nuance: locking a long term contract is not a full answer, because you still cannot predict how much compute you will need. Futures add the optionality fixed contracts cannot.
The moat: breadth, depth, methodology
Kony pressed the obvious question. If establishing a reference price for compute is valuable, what stops copycats from scraping the same data? Carmen welcomes copycats as validation, but is specific about why the work is hard to replicate.
A credible index needs three things. First, breadth. With two or three sources, even real prices can be manipulated; with over a hundred, manipulating all of them at once becomes impractical. Silicon Data pulls from neoclouds, marketplaces, and hyperscalers, taking in hundreds of thousands of data points a day on rental alone. Second, depth. Lumping every H100 into one number is wrong, because H100 machines differ by CPU, RAM, and geolocation. Everything is normalized to a base case first, which requires years of history to do correctly. Third, methodology. A separate index committee governs calibration and restatements, the company is IOSCO compliant, and it goes through third party audit.
Honestly, Silicon Data doesn’t care. We don’t think anything is important unless the data tells us that way.
None of this is invented. It is standard practice for a serious index provider, which is the point: what looks routine to someone from financial services is not routine to a newcomer who has never run an index. That rigor has a cost. Silicon Data started collecting in 2024 but did not publish its first index until 2025, and generally runs three to six months behind a chip’s first trade.
Why compute is not oil
Carmen resists the oil comparison. Compute borrows characteristics from several commodities without matching any cleanly. Like oil it has grades; like electricity it is metered hourly, but unlike electricity it carries residual value. Her sharper distinction is location. An oil oversupply in one US region does not help an electricity shortage in Europe, but compute is more elastic: depending on the workload, you can route to GPU capacity in another geography through an API. Newer NVIDIA architectures are also more stable with less performance variance, pushing the market toward standardization and making indexing more tractable.
The forward curve: reading the market before futures exist
Carmen argues the forward curve matters and has to come before futures. All of Silicon Data’s indices today are on demand rates, but most volume is moving into reserve contracts, and every reserve contract carries an implied forward rate. In principle the curve slopes downward, since committing longer should earn a discount. When it is in contango instead, buyers are paying a premium for long term contracts, expecting prices to rise.
She walks through recent history: last November nearly every curve sloped downward as prices fell; by early this year the H100 curve had flattened while B200 stayed in contango; today she reads the curve as meaningfully more contango than in February.
Doesn’t mean it’s true. It’s a market consensus view today. Doesn’t translate to what will be the future, but it’s very informative.
The contract shift is stark. Last year monthly contracts were common. This year almost no one signs monthly. Even neocloud operators want to lock prices and move to the next site: if their financing term is seven years, a known monthly figure over seven years is ideal. They are chasing consistency, not the highest price.
What this means for lenders
This is where the pricing layer meets credit. Financing a neocloud exposes you to more than the supply and demand curve. You are exposed to residual value, which correlates to future rental rates, and to how well the borrower maintains the hardware.
The due diligence you have to do is more than what’s your configuration. It’s where GPU centers are, how well you maintain, what’s the thermal degradation look like, fraud prevention.
Because you underwrite a specific set of GPUs, not generic exposure, hardware heterogeneity becomes your problem. Kony described a conservative approach on exactly this point: underwrite the cash flow, depreciate residual toward zero, and structure so contracted cash flows repay the balance rather than salvage value. Carmen’s tools fit that model. The forward curve and futures let a lender mark exposure to market daily, gauge whether implied pricing supports a borrower’s uncontracted out years, and hedge term structure risk.
Who trades this, and what comes next
The October CME futures are cash settled and based on on demand neocloud rates. Anyone long GPU on their balance sheet, including financiers, wants to short; consumers who expect to keep renting want to go long. Carmen expects the counterparties to look like the banks that already help AI companies manage FX, now extending to compute cost, plus funds and ETF issuers. Notably, ETFs have already been filed on her CME indices, both long and inverse, before futures even launch, which she reads as strong confidence.
Looking ahead, she expects longer dated curve trading, block trades, physical delivery, and token products, with the honest caveat that GPUs do not trade forever and aging generations must be sunsetted. Her closing pitch: if you need compute resources, Compute Exchange aims to be the NASDAQ for compute; if you need intelligence and benchmarking, think of Silicon Data as the S&P for compute.
Listen to the full episode of Where Capital Meets Compute for the complete conversation.